Show HN: Sentinel – browser agent using 3x+ fewer tokens (open benchmark) (github.com)

A new benchmark introduced by the creator of Sentinel, an AI browser automation framework, reveals significant improvements in token efficiency, reliability, and speed compared to two competitors, Stagehand and browser-use. Conducted using the Gemini 3 Flash Preview model across nine real-world tasks, Sentinel demonstrated remarkable token savings, using 3.13x to 56.93x fewer tokens than browser-use and 1.42x to 13.33x fewer than Stagehand. Notably, Sentinel achieved a perfect reliability rate on all tasks, while Stagehand struggled with multi-step automation flows, particularly failing a critical login/logout task. This advancement is particularly impactful for the AI/ML community as it highlights a shift toward more cost-effective and reliable AI solutions for browser automation. By leveraging Chrome's Accessibility Object Model for task observation, Sentinel addresses common challenges such as high token costs and unstable performance in multi-step processes, making it a compelling option for developers working with automated browser tasks. The detailed methodology and commitment to reproducibility in the benchmarking process provide a valuable resource for further exploration and verification, encouraging wider adoption and potential improvements in AI interactions with web interfaces.